inclusionAI: Ling 3.0 Flash

Ling 3.0 Flash is a text-to-text model from inclusionAI. Its published architecture is a mixture of experts with 124 billion total parameters and approximately 5.1 billion active parameters per token. Select inclusionai/ling-3.0-flash through POST /v1/chat/completions with model and a messages array. Instructions, source material and prior conversation turns belong in the message text.

Use a focused prompt to turn supplied material into a specific deliverable. For a customer reply, provide the customer’s question and the product or policy text the answer must use, then specify the tone and next step. For a document summary, name the sections you need, such as decisions, unresolved questions and actions. For extraction, define the requested fields and how to represent facts absent from the source. For code work, include the relevant function or file excerpt and ask for an explanation, proposed change or review against an explicit requirement.

The published context window is 262,144 tokens, with up to 32,768 completion tokens listed for the model. Budget the source text, instructions, conversation history and requested completion together. Use max_tokens to set the requested output allowance. A document should be supplied as its actual text when you want the answer to depend on its contents; refer to the relevant sections explicitly when a request combines several passages.

The supported parameter list includes tools and tool_choice for function calling. Define the available functions and their argument descriptions in the request. Your application validates a returned call, executes the selected function and supplies its result to the conversation before requesting a follow-up answer. For example, a support workflow can expose a narrowly defined order lookup and use its returned facts when drafting the response. Keep tool results attached to the correct conversation turn so the model can distinguish them from the user’s original request.

The model supports response_format for JSON output, without JSON-schema enforcement. State the intended fields and validate the returned object in your application before using it. Other published controls include temperature, top_p, top_k, min_p, stop, seed and the listed repetition, frequency and presence penalties. The parameter list also includes reasoning and include_reasoning; choose their supported representation from the request specification rather than assuming values from a different edition.

Chat Completions can return a complete response or streamed text. For a complete reply, read the returned message content and retain the response choices and usage. With streaming, assemble the deltas before treating the answer as complete, and retain any returned tool calls with their arguments. The result is text or a function-call request for the next application step; the surrounding application owns execution and the use of the completed answer.